JP3701671B2 - Method and apparatus for testing communication devices using test signals with reduced redundancy - Google Patents

Description

The present invention relates to a communication apparatus test method and apparatus.
When testing a communication device (eg, a communication device such as a telephone line, telephone network, or codec), a test signal is introduced at the input of the communication device and a test is performed on the resulting output of the device. It is known to obtain “objective” test measurements such as signal-to-noise ratio, which can be calculated by an automatic processor. It is also known that a listener listens to the output of a communication device and performs a “subjective” test that expresses opinions on the quality of the output.
Some elements of the communication system are linear. Thus, simple artificial test signals such as discrete frequency sine waves, swept sine wave signals or chirp signals, random or pseudo-random noise signals, or pulses can be applied. The output signal can be analyzed using, for example, Fast Fourier Transform (FFT) or other spectral analysis techniques. One or more such simple test signals can fully characterize a linear system.
On the other hand, the elements included in modern communication systems are increasing in the number of elements that are non-linear and / or change over time. For example, modern low bit rate digital speech codecs that form part of a mobile telephone system have a non-linear response and automatic gain control (AGC), a voice activity detector (VAD) and an associated voice switch, and burst errors are It causes time changes in the communication system in which they form part. Therefore, simple test methods developed for linear systems to obtain an objective measure of communication device distortion or tolerance become increasingly unusable.
Recently, the literature (“Measuring the Quality of Audio Devices” by John G. Beerends and Jan A. Stemerdink, presented at the 90th AES Convention, 19-22 February 1991, Paris in AES Preprints as Preprint 3070 (L-8) by the Audio Engineering Society), using a database of recorded speech as test signals, and a perceptual analysis method designed to respond in several ways to processes that are likely to occur in human hearing It has been proposed to measure the quality of a speech codec for digital mobile radio by analyzing the corresponding output of a codec that uses.
Artificial speech signals (ie, spectrally similar to human speech but with intelligence) along with conventional distortion analysis measurements such as cepstral distance (CD) measurements to measure the characteristics of communication devices. Are also proposed (eg, “Objective Measurement Method for Estimating Speech Quality of Low Bit Rate Speech Coding”, NTT Review, Vol. 1), for example, by Irii, Kurashima, Kitawaki, and Itoh. .3.No.5, September 1991).
When using a test device such as a codec designed to encode human speech and an analysis method based on human hearing, as suggested in the above-mentioned Beernends and Stemerdink document, It is clear to use speech samples. In practice, however, the characteristics of such a test system are not very good.
One artificial voice test signal is described in CCITT Recommendation P50 (Artificial Voices. Vol, Rec. P50, Melbourne 1988, published by CCITT). In the P50 test signal, there are 16 predetermined spectral patterns with a randomly selected sequencer given by segments of a predetermined length, and the transition between the segments is smooth. The P50 signal has long and short term spectral similarity to speech when averaged over about 10 seconds. However, it has been observed that some differences between the P50 test signal and the actual speech are significant when testing the communication system.
Another type of artificial test signal that has been proposed can be found in the literature ("Description and Generation of Sphericaliy Invarient Speech Model Signals" by Signal Processing, Vol. 12 Pt 2, March 1987 by H. Brehm and W. Stammler). Use a spherical non-changing random process (SIRP) as described in. This provides a signal that has the same long-term spectrum as natural speech. The short-term spectrum is not similar to that of natural speech, and the sound grouping over time sounds subjectively very different from natural speech.
A variant called Markov (m) SIRP attempts to model the spectrum for a short period of time, but has a fast random transition between the speech sound that sounds different from the natural speech.
Accordingly, the present invention provides a test signal that includes a sequence of sounds that occur in natural speech with reduced redundancy compared to natural speech.
When using real speech data, you face several problems. To express, there must be a large number of speakers, the speech material from each speaker must be extensive, and the range of voice levels must be considered for each speaker. Therefore, the amount of natural speech material that must be operated by the system being tested to provide reliable characteristics of the system is very large.
The present invention generally has physiological and linguistic constraints that prevent certain combinations of speech sounds from occurring in natural speech and is significant for many of the sounds produced in human speech for testing. It is based on the recognition that redundancy exists. Thus, according to the present invention, speech sounds (either recorded or synthesized) from actual human speech are constructed and implemented in a structure that provides various speech sounds and levels of representation examples without redundancy. Provide a speech-like test signal of possible length. In this way, it is possible to obtain a speech signal (which differs from the SIRP and P50 test signals described above) of a relatively expressive and realistic period, including the natural speech formant structure and temporal structure. .
An analysis method that can be used to analyze the characteristics of a communication system together with the test signal of the present invention is the applicant's international patent application GB 93/01322 published as WO 94/00922, which is hereby incorporated by reference in its entirety. It is described in the book.
With the modification described in the Applicant's European Patent No. 94300073.7 (filed Jan. 6, 1994), which is hereby referred to, the analysis method allows for a set of spectral bands over successive time intervals. An apparatus can be used that periodically derives a set of spectral component signals from a distorted test signal that is responsive to distortion in each. The device is configured to generate a measure of the subjective impact of distortion by the communication device. This measure of subjective influence is calculated to depend on the spread of distortion over time and / or the spectral band.
Other aspects and preferred embodiments of the invention will be apparent from the following description and claims.
The present invention will now be described by way of example only with reference to the accompanying drawings.
FIG. 1 is a block diagram illustrating the structure of one embodiment of the present invention in use.
FIG. 2 is a block diagram showing in more detail the device of one embodiment of the present invention.
FIG. 3 is a block diagram illustrating in more detail a test signal generator that forms part of the embodiment of FIG.
FIG. 4 schematically shows the structure of the test signal over time.
FIG. 5 is a block diagram illustrating the structure of an apparatus for deriving a test signal sequence according to the present invention.
FIG. 6 is a flow diagram illustrating a process performed by the apparatus of FIG.
FIG. 7 is a flow diagram illustrating in more detail a portion of the process of FIG.
FIG. 8 corresponds to FIG. 3 and shows the structure of a test signal generator that forms part of another embodiment of the present invention.
FIG. 9 is a block diagram illustrating in more detail a speech synthesizer that forms part of FIG.
FIG. 10 is a block diagram showing in more detail an analysis device forming part of the embodiment of FIG.
FIG. 11 is a flowchart showing an outline of a process executed by the analysis apparatus of FIG. 8 in the embodiment of the present invention.
[Outline of device]
Referring to FIG. 1, the communication device 1 includes an input port 2 and an output port 3. The test device 4 includes an output port 5 that couples to the input port 2 of the communication device being tested, and an input port 6 that couples to the output port 3 of the communication device being tested.
Referring to FIG. 2, the test apparatus 4 is connected to an output port 5 to supply a speech signal to the test signal generator 7 and a signal received from the communication apparatus 1 is coupled to the input port 6. And a signal analysis device 8 for analysis. As detailed in Applicant's above-referenced international patent application, the analysis device 8 also uses analysis of the test signal generated by the test signal generator 7, which Is shown by the path 9 from the output port 5 to the input port 6 in this embodiment.
The measurement signal output port 10 is also connected to the analysis device 8, where a signal indicative of a certain measure of the tolerance (e.g. distortion) of the communication device is displayed for further processing or on a visual display unit (VDU) not shown. Is output for.
[First Embodiment]
[Test signal generator 7]
As shown in simplified form in FIG. 3, the artificial speech generator is a digital storage device 71 (eg, hard disk or digital audio tape) that contains stored digital data from which the speech signal can be reconstructed. May be included. The stored data may be individual digitized speech samples, which are continuously fed from the storage device 71 to a signal reconstruction means 72 (eg a digital to analog converter (DAC)) connected to the output port 5. Is done. The sample data stored in the storage device 71 includes one or more speech utterances that last for several seconds (for example, about 10 seconds).
A control circuit 73 (eg, a microprocessor) controls the operation of the storage device 71 to select a specific test signal to be output.
Referring to FIG. 4, the test signal data stored in the storage device 71 includes a plurality of segments t.₀, T₁, T₂, ..., t_nIs reconfigured to form a test signal including:
Each segment t₀To t_nCorresponds to different speech sounds (eg different phonemes) or silences.
[Test sequence design]
In this embodiment, the test sequence is pre-generated by a test signal generator 100 (eg, a digital computer), a speech segment storage device 171 is connected to the generator 100, and the storage device 171 actually speaks each phoneme. Store a database of recorded instances of human speech. The sequence generator 100 is also connected to the storage device 71 for writing a test signal sequence to it.
The first starting point of the test signal is a set of 40-50 phonemes, depending on different definitions, such as the International Phonetic Alphabet (see “Mechanisms of Speech Recognition” by WAAinsworth, Pergamon Press, 1976)).
These phonemes are composed of speech and non-speech sounds, first vowels and consonants, third plosives, friction sounds, transition sounds, etc., depending on the spectral content and important temporal structure. It can be divided into several categories.
For the purposes of the present invention, it is not necessary to test a communication device using every single sound produced in natural speech. Instead, it is sufficient to take one or more examples from each of the above categories and form a test signal. Thus, the set of phonemes is reduced to a subset of 10-20 speech segments representing different spectral content (eg, spoken or non-speech sounds) and temporal structure (eg, ending by opening or closing lips), They are selected in step 301 of FIG.
Next, various parameters are considered, especially the statistical variance of level or energy, pitch, duration and formant transition rate. Formants are well known in the speech field to include resonance peaks in the speech spectrum, and their position varies with articulator movement. Each of these parameters is typically characterized by an average value and rarely occurring upper and lower values. In addition, some phonemes and some combinations of phonemes occur on average more frequently than others. Information on these statistical distributions can be found, for example, in the literature (“Telecommunication by Speech” by Richards DL, Butterworths, 1973, or “Speech and Hearing in Communication” by Fletcher H, D. Van Nostrand, 1953, or “Mechanisms of Speech Recognition” by Ainsworth WA, Permagon Press, 1976, or “Vowel consonant and Syllable-A Phonological Definition” by O'Connor JD and Trim JLM, World, Vol. 9, No. 2, August 1953 )It is described in.
The phoneme superset is then a number of instances of each phoneme in the subset, those of normal level (amplitude), those of the high and low extreme values that are usually accepted (eg, 1 standard deviation, 2 standard deviations, Deviated from the average level by some other predetermined statistical amount), as well as for each extreme value of the pitch (high and low) found in natural speech, and for sounds containing formants Constructed in step 302 by taking, for example, those at each extreme value (high and low) of the formant transition rate observed in the speech. Thus, the superset includes a list of as many as 100 speech segments that include selected phoneme instances. The frequency of occurrence values is stored for each phoneme in the superset.
It is recognized that the diversity of speech levels, as measured using the method described in CCITT recommendation P56, shows a standard deviation of approximately 4 dB in the UK telephone network.
Since many of the elements of a communication device (eg echo canceller or voice switch) have time-varying characteristics, the content that each speech segment generates (ie the signal portion that precedes it) is distorted. Is important when determining the degree to which For example, a speech segment preceded by silence or a low amplitude speech segment may be clipped by a voice switch, while such clipping will not occur if it is preceded by a high amplitude speech segment.
Thus, when constructing a test signal from a superset of speech segments generated in this way, the same speech segment is represented in the test signal in more than one content (ie preceded by different speech segments). Care is taken to ensure or the test signal is observed in another way to ensure that it contains many different types of transitions from one type of speech segment to another. A “silence” segment is also included in the superset because a short silence occurs during the speech.
However, linguistic reasons for limiting the length of a series of consonants, for example, due to physiological limitations that prevent the articulatory organ from moving from one location to another immediately in the human speech system Some transitions from one phoneme to another for any of them (which can vary between languages) do not occur in natural speech. Thus, according to the present invention, several sequences of speech segments are not allowed when constructing the test signal.
FIG. 7 shows an example of the algorithm used in step 303 to generate a test sequence. The sequence lasts about 180 seconds and consists of a series of segments, which are preferably variable but typically each period on the order of a few hundred milliseconds.
In step 201, the first speech segment (t in FIG.₀, Index i = 0) is selected from a superset of speech segments on an arbitrary or random (pseudo-random) basis.
In step 202, the speech segment has already been used more than a predetermined number of times in the sequence, or the speech segment is the same as the previous speech segment (t_i-1) Is tested to see if it has already occurred in a later sequence. If any of these criteria are met, the selected speech segment is not included in the sequence, and instead the selection step 201 is repeated to find a new speech segment.
In step 203, the selected segment is the previous segment (t_i-1) Physiological or linguistic rules are applied to determine whether it is possible to follow. For example, if a segment is a consonant that occurs after two preceding consonants, or a friction sound that occurs after two preceding consonants, it is rejected. Such generation rules are well known, for example, in the synthesis of text and speech. If the selected segment does not meet these generation rules, step 201 is repeated again.
In step 204, the length of the sequence is tested, and if the sequence already consists of a predetermined number of segments (ie i = N, predetermined consonant), the sequence ends and the process proceeds as shown in FIG. Proceed to step 304.
In step 205, the number of speech segments from the last silence segment in the sequence is tested. Silence segments are members of a superset of segments and are therefore selected on a random basis in step 201. However, if a speech segment longer than 2 seconds (for example) has occurred without a silence segment, the silence segment is selected at step 206 in preference to the segment selection at step 201.
Otherwise, if the segment selected in step 201 is added to the sequence in step 207, the index counter i is incremented in step 208.
When a list of speech segments that limit the sequence is selected, in this embodiment, a sample sequence of signal values recorded in the storage device 71 reads the corresponding stored speech segment from the storage device 171 and the two segments do not touch. This is obtained by digitally editing the sequence at the transition between speech segments so that the sequence is removed at step 304.
The joining or concatenation of speech segments in step 304 can be performed using overlap and additive techniques used to join diphones in the synthesized announcement. In literature (Speech Communicaiton by E.Moulines and F.Charpentier, Vol.9, No.5 / 6, 453, or proceedings IEEE, International Conference on Acoustics, Speech and Signal Processing (ICASSP) 1989 by C.Hamon et al.) The described pitch-synchronized overlap and additive technique (PSOLA) is preferred.
In this technique, in the edge region of two speech segments to be joined, if both are speech segments, the different pitches of the two segments are gradually changed towards the intermediate pitch at the joint of both, The waveforms of the two speech segments are added redundantly as described in more detail in the above document.
Any total energy mismatch between adjacent speech segments is corrected by linear rescaling of the waveform over the overlapping region, and spectral mismatch is described in, for example, the literature (IEEE proceedings of International Conference by F. Charpentier and E. Moulines). on Acoustics, Speech and Signal Processing (ICASSP) 1989).
Thus, the sequence generator 100 concatenates selected segments from the segment store 171 to progressively write a continuous test signal sequence of digital signal samples to the signal store 71, followed by digital-to-analog conversion. Used in the test signal generator 7 to reconstruct the speech signal via the generator 72.
[Second Embodiment]
Referring to FIG. 8, in this embodiment, the test signal generator 7 is controlled by a synthesis parameter supplied by a parameter storage 371 to generate a synthesized speech signal via a digital-to-analog converter 72. Includes synthesizer 370.
The speech synthesizer 370 is described in the literature (Journal of the Acoustic Society of America, Vol. 67, No. 3, March 1980 by DH Klatt), and its structure is shown in FIG. Preferably, it is a cascaded / parallel formant synthesizer that can be constituted by a digital signal processing device (DSP). It is essentially a parallel bank and series cascade of resonant filters 371 connected to the output of an impulse generator 373 (generating non-speech speech) and a random number generator 374 (generating non-speech speech) The filter 372 is configured. The repetition rate and filter characteristics of the impulse generator 373 are control parameters supplied from the parameter storage 371.
In this embodiment, the sequence generator 100 operates in the same way as in the previous embodiment, except that the sequence storage device 171 stores a sequence of control parameters that allows speech segment reconstruction. The transition is easily accomplished by tapering the values of each synthesizer parameter in adjacent speech segments to each other for an overlap period.
Instead, the type of speech synthesizer 370 or stored announcement used in a synthesizer that synthesizes text and speech may be used. Such a device operates using several rules that can partially replace step 203 of FIG. 7 to connect adjacent phonemes without further input. In this case, the storage device 71 stores and outputs only phoneme identifier data, level data, pitch and period or speed data.
[Analyzer 8]
Referring to FIG. 10, the analysis device 8 receives a signal from the input port 6 and receives an analog to digital converter (ADC) 81 configured to generate a corresponding digital pulse train and a digital output of the ADC 81. A combined arithmetic processor 82 (eg, a microprocessor such as an Intel 486 processor or a digital signal processor such as a DSP32C from Western Electric or a TMS C30 device from Texas Instruments) and a processor 82 It includes a storage device 83 that stores instruction sequences and provides an operational memory for storing operation results, and an output line 84 from the processor 82 connected to the output 10.
The processes performed by the processor 82 correspond to those described in the applicant's above-referenced international patent application WO 94/00922 and European patent application 94300073.7, both The book is hereby a reference. However, according to the present invention, the data used when obtaining the test signal can also be used to assist the analysis performed by the analysis device 8.
In the applicant's above-referenced patent application, the analysis process calculates one or more measures of distortion of the test signal over some time interval. In the present invention, the time interval corresponds to the time segment in which the analysis is performed, and the distortion measure calculated for all speech segments affects one or more global distortion measures, where each speech segment distortion has a scale. The amount that affects the overall measurement is determined according to the frequency of occurrence of the speech segment. Thus, infrequently occurring phonemes, infrequently occurring levels, infrequently occurring pitches, infrequently occurring periods, and infrequently occurring content may all be present in the test signal. The distortions they receive are given a suitably low weight when determining the characteristics of the communication device.
Referring to FIG. 11, according to one embodiment, the test signal generated by the generator 7 in operation is provided to the analyzer 8 at line 9 at step 101. In step 102, the sequence of weighted data, each segment t of the test sequence_iOne weight value W for_iIs supplied from the test signal generator 7 to the analysis device 8. The weighted data is obviously equal to the probability of occurrence of a speech segment involved in natural speech, i.e. its level, its pitch, and the probability of occurrence of a phoneme at that period, and optionally the probability of that phoneme following the previous speech segment. You may consider it. Weight W_i(I = 0 to N) is stored in the storage device 71 together with the test signal.
In step 106, the test signal generator 7 generates a test signal supplied to the analysis device 8 via the communication device 1. In step 109, the analysis device 8 preferably uses a perceptual model, and very preferably uses the method described in the above-referenced patent application specification (any example thereof). A measure of distortion is then calculated between the undegraded test signal supplied on line 9 for each segment and the test signal distorted by the communication device 1. The distortion measure for each segment may be, for example, error loudness or the total amount of errors or distribution of errors described in the patent application. Each segment E_iIf a distortion measure (eg, its segment and / or the amount or loudness of distortion in the spectrum and its temporal spread) is specified, then the overall distortion measure D of the communication device 1._iIs the segment distortion measure E_iIs calculated in step 210 by taking the weighted sum of. In other words,
D = ΣE_iW_i
This measure of the characteristic is output in step 111 to, for example, a visual display device (not shown).
In another embodiment, weighted W_iRather than the test signal generator 7 supplying the analyzer 8, the analyzer 8 recognizes the clean test signal on line 9 by matching the test signal to stored data representing a complete set of phonemes. A weighting factor W associated with the phoneme data stored in accordance with the derived phoneme identifier, level, etc._iA speech recognizer configured to use
[Results of Invention]
The efficiency of the test signal of the present invention was compared with that of several known prior art test signals.
Compared to the test signal generated by SIRP, the test signal of the present invention provides a signal having a structure over time very close to that of normal human speech. This has been found to be very effective when evaluating an actual communication system in which the effects of time-dependent elements such as voice switches are subjectively very significant.
Furthermore, since the present invention includes a spoken portion (as opposed to a SIRP speech signal), it can better characterize spectral distortions of important formant frequencies that are subjectively significant to the listener.
Compared to the CCITT Recommendation P50 signal, the present invention provides a highly altered speech test signal that includes features such as uttered frictional sound that are not present in the P50 signal. In addition, the present invention provides a test signal with a clear formant structure that is not present in the P50 signal. This is important because the sliding or moving formant frequencies (especially those of the second formant) are important in phoneme recognition and thus important for the clarity of speech signals in communication systems.
Finally, the time structure of the speech signal of the present invention is more variable than that of the P50 test signal, includes fast transitions, onsets and stops, and thus has a detailed amplitude structure over time that is close to natural speech. Due to the time dependence of the elements of the communication network, the relatively slowly changing P50 test signal does not emphasize distortion that clips fast transitions.
Compared to actual human speech, the test signal of the present invention greatly reduces redundancy, and it occurs excessively or infrequently without leaving the device unserviced for a significantly longer period of time. The situation that does not occur can be applied to the communication device 1.
In the present invention, a small representational subset of speech segments (selected from dozens of known music) is used to construct a test signal from these sounds assembled into sequences of different content. Since distortion is measured, the test sequences are relatively different from each other, or more generally, a series of sounds that are prone to distortion when one follows another. It is even more important to include.
[Other alternatives and modifications]
It will be apparent from the foregoing description that numerous modifications can be made to the embodiments described above without altering the operating principles of the invention. For example, the signal from the output port 5 may be supplied in digital form to the input port 2 of the communication device and the ADC 81 may be removed. Instead, an electromechanical transducer is provided at the output port 5 and the signal can be supplied as an audio signal. In the latter case, the test signal is CCITT P.51 (Recommendation on Artificial Ear and Artificial Mouth, Vol. 5, Rec P. 51, Melbourne 1988) and the above mentioned UK patent application GB2218300, both of which are hereby incorporated by reference. (8730347) may be supplied via an artificial mouse described in the specification.
Although the connection to an actual communication device has been described above, it is equally possible to program the computing device to simulate the distortion introduced by the communication device. This is because it is relatively easy to characterize a large number of such distortions (eg, by VAD or codec). Thus, the present invention applies equally to embodiments in which signals are supplied to such a simulation device, and the output of the communication device with simulated distortion is processed. In this way, the tolerance for a listener of a number of complex and non-linear communication device combinations can be modeled before assembling or connecting such devices in the field.
Although the analyzer 8 and the test signal generator 7 are described as separate hardware, they can actually be implemented by a single properly processed digital processor and are also shown in the above example. The communication device simulator can be constituted by the same processor.
In this description, “phoneme” refers to a sound whose speech content has been modified in its normal usage, but for convenience the term “phoneme” is used to denote a single repeatable human speech sound. ing.
While the embodiments shown above relate to communication device testing, application of the new aspects of the present invention to other testing or analysis is not excluded.
Accordingly, protection is sought against the new matter or combination of new matters described herein and variations thereof, whether such matter or variation is within the scope of the following claims. Will be apparent to those skilled in the art.
Instead, the storage device 71 stores the speech data in the form of filter coefficients and drives it, for example, an LPC speech synthesizer, or stores it in the form of high level data (eg phoneme, pitch and identifier data). Thus, the phoneme synthesizer including the reconstruction unit may be driven.

Claims (13)

A method of testing a communication device by supplying to the communication device a test signal that consists of a sequence of sounds that occur in natural speech,
Wherein the statistics generation of test signals, to test you are, communication device different from the generation of statistics on natural speech as sound hardly occurs in natural speech is contained at a high incidence. A test signal for testing a communication device,
Nature comprises a sequence of sounds generated in the speech test signal that has been reduced redundancy compared to natural speech. A test signal for testing a communication device ,
Comprises a sequence of speech sound corresponding to the part of the word found in natural speech,
Some, test signals that are present in the plurality of cases have characteristics different from the value of the speech in the sequence of the speech sound. 4. The signal according to claim 3, wherein one feature value is an average value, and at least the other one is a feature value that occurs rarely. The signal according to claim 3 or 4, wherein the characteristic is composed of one or more of volume, pitch, period, and formant transition rate. A test signal for testing a communication device,
It said test signal comprises a sequence of speech sound corresponding to the part of the word found in natural speech,
It said sequence is intended to represent a speech sound generated by the natural speech, including the transition between the at least one speech sound and speech sound adjacent to it, test signal. 7. A signal as claimed in any one of claims 2 to 6 including a concatenated segment of pre-recorded human speech. 7. A signal as claimed in any one of claims 2 to 6 comprising a synthesized signal generated by a parametric speech synthesizer. 9. A signal according to any one of claims 2 to 8, wherein some of the sounds are present in a plurality of instances that immediately follow a plurality of different preceding sounds. A method for testing a communication device, comprising:
Supplying a communication device with a test signal including a plurality of segments corresponding to speech-like sounds having a high relative frequency of sounds that occur relatively rarely in natural speech;
Analyzing the signal output from the communication device for each sound,
A method of testing a communication device thereby forming an overall measure of the distortion produced by the communication device based on each sound distortion and its frequency of occurrence in natural speech. Previously generated and data storage means configured to store the recorded Ru test signal sequence, a test signal according to any one of the stored sequences the reconstructed section with a comprise that claims 2 to 9 Device to do. The prerecorded Ru sequence device according to claim 11 which contains a segment of natural speech that will be recorded. 12. The apparatus of claim 11 including a speech synthesizer and data storage means for controlling the speech synthesizer to generate the test signal sequence.

Images